Experts say today’s chatbots do not add much to bioweapon risk

In short: Researchers and policy groups largely agree that today’s general chatbots do not meaningfully help someone create a deadly pathogen, but more biology-focused AI may raise bigger safety concerns.

What's going on

Several recent reviews take a split view of AI and biological risk. They say general-purpose chatbots, the kind many people use for writing help or quick answers, mainly summarize information that is already public. The U.S. National Security Commission on Emerging Biotechnology said in 2024 that these systems do not significantly increase bioweapon risk because they do not add much beyond what someone could find online.

A separate 2024 evidence synthesis reached a similar conclusion. It argued that today’s publicly available large language models (LLMs, AI systems trained to predict and generate text) do not meaningfully increase biological risk compared with ordinary internet access.

At the same time, analysts including RAND have drawn a line between general chatbots and more specialized biology models. These biology-focused systems can work more like a design assistant, not just a librarian. Instead of explaining biology, they may help design or improve biological sequences (like giving suggested “letters” in DNA) and predict traits such as how easily something could spread or evade immune defenses.

What to watch

The main concern is capability level. As models trained on biological data become more helpful for real lab work, experts argue they should be tested and restricted before wide release, similar to checking a tool’s safety before selling it broadly. OpenAI has also warned that future biology-capable models could be misused by people with little expertise, and it says it is preparing safeguards for those future capabilities.

Source: NYTimes